Skill creator

Skill Creator

Create new Claude Code skills interactively, following all conventions from SKILLS_GUIDE.md and the learning/configurability/feedback patterns proven in the best existing skills.

Preferences

On startup, use the Read tool to load ~/.claude/skills/skill-creator/preferences.md. If missing, treat as "no preferences set".

Context

Before starting:

• Use the Glob tool to discover existing skills at ~/.claude/skills/*/SKILL.md (so you don't reinvent or collide with a name).
• Read ~/.claude/skills/SKILLS_GUIDE.md once at Step 3 — it's the canonical manifest.
• Treat svg-art, should-i-buy, and step-through as the gold-standard reference for learning + feedback patterns.

Command routing

Check $ARGUMENTS:

• help → display help, then stop
• config → interactive setup, then stop
• reset → delete ~/.claude/skills/skill-creator/preferences.md, confirm, stop
• --from-description "..." → skip Q1, use the quoted text as the skill's purpose
• anything else (including empty) → run the skill

Help

Skill Creator — Creates Claude Code skills, manifest-compliant by default

Usage:
  /skill-creator                             Interactive skill creation
  /skill-creator <name>                      Start with a name pre-filled
  /skill-creator --from-description "..."    Start from a plain-text description
  /skill-creator config                      Set preferences
  /skill-creator reset                       Clear preferences
  /skill-creator help                        This help

Examples:
  /skill-creator                             Full interactive walkthrough
  /skill-creator deploy-preview              Create a skill named "deploy-preview"
  /skill-creator --from-description "A skill that triages PR comments"

Current preferences:
  (loaded from ~/.claude/skills/skill-creator/preferences.md)

Config

Use AskUserQuestion to collect:

• Q1 — Default skill location: ~/.claude/skills/ (standard) or custom path
• Q2 — Auto-update inventory in SKILLS_GUIDE.md after creation: yes/no
• Q3 — Publish repo path (e.g. ~/Dev/claude-skills) or "none"
• Q4 — Default learning posture: learning-on (include feedback + journal + sessions) / learning-off (stateless tool by default) / ask-each-time

Save to ~/.claude/skills/skill-creator/preferences.md.

Reset

Delete ~/.claude/skills/skill-creator/preferences.md and confirm: Preferences cleared. Using defaults.

First-time detection

If no preferences file exists:

First time using /skill-creator? Run /skill-creator config to set defaults, or just continue with sensible defaults.

Then proceed.

---

Manifest patterns reference (read once, apply every time)

Every generated SKILL.md must satisfy these — no exceptions unless explicitly justified to the user.

Required (from SKILLS_GUIDE.md)

1. Frontmatter — name (kebab-case, ≤64 chars), description (third-person, "Use when …" trigger, ≤1024 chars), argument-hint, minimal allowed-tools.
2. disable-model-invocation: true — required for any skill that creates / modifies / pushes / posts.
3. AskUserQuestion in allowed-tools — every skill needs it for help/config/reset.
4. Subcommand pattern — help, config, reset, default behaviour.
5. Read of preferences on startup — never ! backtick interpolation for ~/.claude/ paths.
6. First-time detection — warm one-liner, never blocking.
7. No ! backtick interpolation — use runtime "On startup, use Bash to …" instructions instead.
8. Under 500 lines — push reference material to reference/ or examples/.
9. Numbered workflow — checklists, not prose.
10. Graceful degradation — every external read/tool call must have a fallback.
11. Validation loops — after any action, verify before moving on.

Required for skills with discrete sessions / outputs (default ON unless stateless)

12. Three-tier preferences format:

    # /<skill> preferences
    Updated: YYYY-MM-DD
    
    ## Defaults
    - knob: value
    
    ## Profile (optional — edit freely)
    - user-editable bias
    
    ## Learned
    - (populated from feedback over time)
    

13. feedback subcommand — /skill feedback collects ratings on the most recent session via AskUserQuestion.
14. feedback-journal.md — append a per-session block with a Signal: one-liner generalisation.
15. sessions/YYYY-MM-DD-HHMM.md — incremental session log when the skill has discrete invocations worth tracking.
16. Promotion rule — when 3+ sessions show the same signal, promote it to ## Learned in preferences.md. Mention once: "Noticed you consistently … Saved as standing default."
17. Drift correction — when a Learned rule is contradicted in 2 newer sessions, demote it (log the demotion in the journal). Never leave stale rules.
18. Step 0 — Load learning context — read preferences.md and feedback-journal.md before the main workflow. Continue silently on missing files.
19. Invite feedback line at the end of the workflow: Run /<skill> feedback — even one rating helps me sharpen defaults.
20. Reset clears the lot — preferences.md, feedback-journal.md, sessions/, any resume-state.md.

Required for skills that render UI (HTML, React, SVG, posters, dashboards, reports)

26. DESIGN.md is mandatory. Any skill that generates user-facing UI — HTML reports, React components, SVG artefacts, Markdown with structured visuals, posters, slide decks — must include a DESIGN.md at the skill root. The DESIGN.md is the renderer contract: it specifies aesthetic direction, type system, colour tokens, layout, visualisation principles, anti-patterns to avoid, and accessibility guarantees. Renderers reach for DESIGN.md before generating any markup, and prefer its tokens over hard-coded values.
27. At runtime, the skill MUST Read DESIGN.md before rendering any UI — and the workflow must explicitly cite it (e.g. "Step N — Load DESIGN.md. Apply its tokens, type system, anti-patterns, and a11y contract to every emitted artefact."). Treat missing DESIGN.md the same as missing preferences: degrade gracefully with built-in defaults, but never silently ignore it when present.
28. DESIGN.md sections (canonical structure):

    # /<skill> design
    
    ## Aesthetic direction
    The committed point of view. References, tone, what this is NOT.
    
    ## Type system
    Display font + body font + mono font, with fallbacks. Scale, weights, letter-spacing.
    
    ## Colour tokens
    CSS custom properties. AA-compliant for any token used as text. Dark-mode mirrors.
    
    ## Layout
    Max widths, spatial moments, grid/asymmetry rules.
    
    ## Visualisation principles
    How charts encode data. Density binning, sparse-data fallback, redundancy beyond colour.
    
    ## Anti-patterns to avoid
    Project-specific don'ts, plus the universal slop list.
    
    ## Renderer contract
    Hard guarantees: ARIA on every SVG; contrast ≥4.5:1 for text; no color-only encoding; no
    `title`-only tooltips; pre-computed colour-mix fallbacks; anonymisation rules; etc.
    
    ## Component patterns
    The reusable HTML/JSX/SVG snippets — each shown inline, using tokens not hex values.
    

29. DESIGN.md vs preferences.md — DESIGN.md is the skill's aesthetic contract (versioned in the skill, identical for every user). preferences.md is the user's runtime configuration (per-machine, edited freely). When they conflict, user preferences win for runtime knobs (e.g. tone, depth, framework); DESIGN.md wins for renderer guarantees (e.g. a11y contract, anti-patterns). Encode this priority in the workflow: load DESIGN.md, then preferences, and only let preferences override DESIGN.md tokens flagged as user-tunable.
30. Anti-pattern baseline. Every UI-rendering skill must include the universal slop list in its DESIGN.md's "Anti-patterns" section (with [[from skill-creator]] link):
    • Inter / Roboto / Arial as primary font choice — pick distinctive pairings
    • Purple gradients on white
    • Glassmorphism / backdrop-blur for decoration
    • Hero metrics without scale legend
    • Color-only encoding (must pair with shape/size/position/texture)
    • Tooltips via title="" only (need focusable + visible alternatives)
    • Nested cards (cards inside cards)
    • Bounce / spring easing on serious / professional content
    • font-feature-settings: "ss01", "cv11" applied to non-Inter fallback fonts (dead code)
    • -webkit-font-smoothing: antialiased (harms low-vision users)

Recommended (apply when relevant)

21. resume-state.md + /<skill> resume — for any skill where the user can stop mid-flow.
22. Confirmation policy — destructive-only by default; configurable to "every action" via preferences.
23. Pre-selection logic in AskUserQuestion — bias the default option using Learned rules.
24. Configurability ≥ 3 knobs — depth, verbosity, output format, source selection, confirmation policy. A skill with only 1 knob is under-specified.
25. Templates (templates/<name>.md) — when a skill has a recurring "preset workflow" with parameterized inputs (e.g. /publish-note --from-template=raw-idea "<idea>"), expose it as a template file inside the skill directory. Each template is a YAML-frontmatter + markdown file declaring inputs, tasks, constraints, tools, postProcesses. The skill should support --from-template=<name> routing in addition to its standard flow. See ~/.claude/skills/publish-note/templates/raw-idea.md for the canonical example. Templates are how a skill compresses an entire conversational workflow ("I want X, with these constraints, ending in Y") into a single command.

When to skip the learning tier (12–20)

Skip only if all of the following are true:

• The skill has no per-session output worth rating (e.g. aws-mfa writes a credential file — there's nothing to thumbs-up).
• The skill is purely procedural and identical every run.
• There's no per-severity / per-output category that could be biased over time.

If skipping, document it in the SKILL.md's Principles section: Stateless by design — no feedback loop because [reason].

---

Step 1 — Understand the skill's purpose

If $ARGUMENTS contains --from-description, use that text. If $ARGUMENTS is a bare name, pre-fill the name and ask the rest. Otherwise, ask:

What should this skill do? Describe the workflow in a sentence or two.

Extract:

• Working name (kebab-case)
• Core verb (creates, analyses, fetches, organises, walks through, …)
• Trigger phrase (when should the user reach for it)

Confirm with the user: I'll call this /skill-name — a skill that [one-liner]. Sound right?

Step 2 — Determine side effects, context, and learning posture

Use AskUserQuestion with 3 questions in one batch:

Q1 — Side effects: Does this skill create, modify, push, or post anything?

• Yes → disable-model-invocation: true
• No (read-only) → omit disable-model-invocation
• Pre-select Yes if the description mentions create/post/push/update/write/deploy.

Q2 — Context needs: Does this skill need the current conversation history?

• Yes (builds on what we've discussed) → no context field
• No (self-contained, works from its own inputs) → context: fork
• Pre-select fork if the skill processes external input only (URLs, file paths, pasted text).

Q3 — Learning posture: Should this skill learn from per-session feedback?

• Yes — full learning tier (preferences + journal + sessions + feedback subcommand) — pre-select for any skill with discrete outputs, decisions, or judgement calls
• No — stateless tool — pre-select only for purely procedural skills (e.g. credential refresh, fixed pipeline runners)
• If Yes, items 12–20 of the manifest reference are now mandatory.

Step 3 — Identify tools needed

Read ~/.claude/skills/SKILLS_GUIDE.md to confirm conventions.

Use AskUserQuestion (multiSelect):

Which capabilities does this skill need?

• Read files → Read, Glob, Grep
• Edit files → Write, Edit
• Run commands → Bash (use Bash(git *) / Bash(gh *) glob patterns when scoping)
• Ask questions → AskUserQuestion (always include — auto-add even if not picked)
• Web access → WebSearch, WebFetch
• Linear → use mcp__claude_ai_Linear__* (NOT mcp__linear-server__*):
  • Read: list_issues, get_issue, list_comments, list_projects, get_project, list_teams
  • Write: above + create_issue, update_issue, create_comment
• Vercel / Slack / Notion / Chrome / Playwright → pick specific MCP tools by action

Always include AskUserQuestion. Always keep the list minimal — granting Write/Edit to a read-only skill is a bug.

Step 4 — Design the workflow

Apply the manifest reference. Draft a numbered workflow that includes (in order, when learning tier is on):

1. Step 0 — Load learning context (read preferences + journal)
2. Domain-specific steps (4–8 of them)
3. Step N–1 — Final summary
4. Step N — Invite feedback (one-line pointer to /<skill> feedback)

For stateless skills, drop steps 0 and N.

Identify ≥3 configurability knobs (depth, verbosity, format, source, confirmation policy, …). Identify the 1–3 decision points that get an AskUserQuestion. Identify what the feedback subcommand will rate (per-severity defaults? per-output style? per-source pace?).

Present the outline:

Here's the workflow:

1. Load learning context
2. … N. Invite feedback

Configurability knobs: [list]. Feedback dimensions: [list]. Want to adjust?

Step 5 — Determine arguments and flags

Propose:

• Positional argument (if any)
• Standard flags: --filter, --start, --from, --dry-run, --verbose (only those that fit)
• Subcommands beyond the standard four: feedback, resume (if applicable)

Confirm via AskUserQuestion or inline.

Step 5.4 — DESIGN.md (ask if the skill renders UI)

If Step 2 / 3 / 4 surfaced that the skill emits user-facing UI (HTML files, React components, SVG artefacts, slide decks, posters, structured-visual markdown), ask once via AskUserQuestion:

This skill renders UI. Scaffold a DESIGN.md design contract?

• Yes — canonical structure (recommended) — adds DESIGN.md with the eight standard sections (aesthetic direction, type system, colour tokens, layout, visualisation principles, anti-patterns, renderer contract, component patterns). Workflow gets a "Load DESIGN.md before rendering" step. Renderer references tokens, not hex values.
• Yes — minimal stub — DESIGN.md is created with section headings only; user fills in.
• No — skip. (Acceptable only if the skill emits raw text / data, no styling decisions.)

If Yes:

1. Create ~/.claude/skills/<skill-name>/DESIGN.md with the canonical structure (see manifest item 28 and the example at ~/.claude/skills/emotional-recap/DESIGN.md).
2. Add to the skill's ## Context section: On startup, use the Read tool to load DESIGN.md. Apply its tokens, type system, anti-patterns, and renderer contract to every emitted artefact. If missing, fall back to built-in defaults but flag once.
3. Add a workflow step before the render step: ### Step N — Load DESIGN.md.
4. In the structure checklist (Step 6), tick the "DESIGN.md loaded before any render" box.
5. Seed the Anti-patterns section with the universal slop list from manifest item 30.

DESIGN.md is the skill's renderer contract: aesthetic + accessibility + token guarantees, versioned in the skill repo. It is distinct from preferences.md (which is per-user runtime configuration). When they conflict, DESIGN.md wins for renderer guarantees (a11y, anti-patterns); preferences wins for user-tunable knobs (tone, depth, framework choice).

Step 5.5 — Templates (optional, ask once)

Ask once via AskUserQuestion:

Will this skill have parameterized preset workflows (templates)?

• Yes — I expect users to invoke this with a recurring pattern of inputs/constraints (e.g. "raw idea → finished output with these tools and constraints"). Add templates/<name>.md directory + --from-template=<name> routing.
• No — single workflow only.
• Maybe later — leave the directory off; can be added later via /skill-templates (planned skill).

If Yes:

1. Create ~/.claude/skills/<skill-name>/templates/ directory.
2. For each template the user names, scaffold templates/<name>.md with the canonical frontmatter:

   ---
   name: <template-name>
   description: >-
     <one-liner — what this template does, what it skips, what it adds>
   inputs:
     <input>:
       type: string | list
       required: true | false
       default: <value>  # if not required
   tasks: [<ordered list of step keywords>]
   constraints: [<rules the run must respect>]
   tools: [<preferred tools>]
   postProcesses: [<actions after main work>]
   ---
   
   # Template: <name>
   
   ## When to use
   <trigger scenario>
   
   ## Flow (default tasks)
   1. **<task>** — <how to execute>
   ...
   
   ## Default constraints (why each one)
   - **<constraint>** — <reason>
   
   ## Examples (from prior runs)
   - <date>: <what happened>
   

3. Add templates and --from-template=<name> routes to the skill's Command routing block.
4. Add a ## Templates section in the SKILL.md explaining the convention and how the skill loads/parses templates.

Reference: ~/.claude/skills/publish-note/templates/raw-idea.md is the canonical example.

Step 6 — Generate the SKILL.md

Assemble. Use the structure checklist below. Walk through it top to bottom — every box must be ticked or explicitly justified as "skip" with a Principles note.

Structure checklist

Frontmatter

• name — kebab-case, matches directory
• description — third-person, includes "Use when …" trigger, ≤1024 chars, mentions learning if applicable
• argument-hint — reflects positional + flags + subcommands
• disable-model-invocation: true — set if side effects confirmed in Step 2
• context: fork — set if confirmed self-contained in Step 2
• allowed-tools — minimal list from Step 3, includes AskUserQuestion

Top-level sections

• ## Preferences — runtime Read instruction for preferences.md + Defaults list
• ## Context — runtime Bash/Read instructions, no ! backticks
• ## Command routing — help / config / reset / feedback (if learning) / resume (if applicable) / default
• ### Help block — CLI format with Current preferences line
• ### Config block — AskUserQuestion + write to preferences.md in three-tier format
• ### Reset block — deletes preferences + journal + sessions + resume-state when learning tier is on
• ## First-time detection — warm one-liner, non-blocking

Workflow

• ### Step 0 — Load learning context (skip only for stateless)
• Numbered domain steps (4–8)
• ### Step N — Final summary with one follow-up AskUserQuestion
• ### Step N+1 — Invite feedback line (skip only for stateless)

Learning tier (when on)

• ## Feedback & learning section with: feedback subcommand handler, journal append format with Signal:, promotion rule (3+ sessions), drift correction (2 contradictions → demote)
• Pre-selection in domain AskUserQuestions biased by Learned rules

DESIGN.md (when Step 5.4 = Yes — i.e. skill renders UI)

• DESIGN.md exists at skill root with the 8 canonical sections (aesthetic / type / colour / layout / visualisation / anti-patterns / renderer contract / component patterns)
• Anti-patterns section seeded with the universal slop list (manifest item 30)
• Renderer contract section enumerates: ARIA on every SVG, contrast ≥4.5:1 for text, no color-only encoding, no title-only tooltips, pre-computed color-mix fallbacks, anonymisation rules (if user content is rendered)
• ## Context of SKILL.md instructs runtime to Read DESIGN.md before any render
• At least one workflow step explicitly references DESIGN.md tokens ("apply tokens from DESIGN.md")
• Renderer / report-structure references DESIGN.md tokens rather than hard-coded hex values

Templates (when Step 5.5 = Yes)

• templates/ directory exists with at least one scaffolded template
• Each templates/<name>.md has the canonical frontmatter (inputs, tasks, constraints, tools, postProcesses)
• ## Templates section in SKILL.md documents loading + parsing
• Command routing handles templates (list) and --from-template=<name> (run)

Closing

• ## Principles — 5–8 rules. Always include: "Graceful degradation on missing learning state" and "Destructive actions still need confirmation" if any side effects
• Total ≤500 lines (push reference material to reference/ if over)

Dynamic context injection

Never use !`...` backtick interpolation in personal skills. Use runtime instructions:

## Context
_On startup, use Bash to detect: current git branch, git status, repo name. Skip any that fail._

Preferences file format (mandatory three-tier when learning tier is on)

# /<skill-name> preferences
Updated: YYYY-MM-DD

## Defaults
- knob: value
- knob: value

## Profile (optional — edit freely)
- user-editable bias lines

## Learned
- (populated from feedback over time)

Feedback subcommand template (when learning tier is on)

## Feedback & learning

When invoked as `/<skill> feedback`:

1. Find the most recent `sessions/YYYY-MM-DD-HHMM.md`. If none, say `No recent session found.` and stop.
2. Print a one-line summary of that session.
3. Ask via `AskUserQuestion` (3–4 questions in one batch). Tailor questions to the skill's output dimensions.
4. Append to `~/.claude/skills/<skill>/feedback-journal.md`:

{session slug} — {date}

• {dimension}: {answer}
• Signal: {one-line generalisation}

5. Promotion rule: 3+ sessions same signal → promote to `## Learned` in `preferences.md`. Mention once.
6. Drift correction: 2 contradictions on a Learned rule → demote, log in journal.

Principles starter set

Include these by default; cull or rewrite as needed:

• Graceful degradation — continue with built-in defaults if preferences.md / feedback-journal.md can't be read
• Learn quietly — promote a rule only after 3+ consistent signals; mention once when promoting
• No fabricated structure — if input is ambiguous, ask rather than invent
• Destructive actions need confirmation — auto mode does not bypass file deletes / pushes / external posts
• Stop means stop — when user halts, save resume state if applicable, do not continue in-flight work
• Skill orchestrates, does not re-judge — preserve user inputs verbatim where applicable

Step 7 — Write the files

1. Create ~/.claude/skills/<skill-name>/.
2. Write SKILL.md.
3. If reference material exceeds 100 lines, split into reference/ files.
4. Do not pre-create preferences.md, feedback-journal.md, or sessions/ — those are created on first use.

Step 8 — Self-audit (mandatory)

After writing, re-read the generated SKILL.md and tick every box in the Structure checklist (Step 6). For each box that is unticked, either:

• Edit the SKILL.md to fix it, or
• Add a one-line note in the Principles section justifying the skip (e.g. Stateless by design — no feedback loop because [reason]).

Then run a quick line-count check via Bash. If over 500 lines, propose a split into reference/.

Print the audit result to the user:

Manifest audit: 23/24 boxes ticked.
Skipped: feedback subcommand — stateless by design (per Step 2 Q3).
File: ~/.claude/skills/<skill-name>/SKILL.md (LINE_COUNT lines)

Step 9 — Post-creation actions

Present options via AskUserQuestion:

• Test it — Run /<skill-name> help to verify the help output
• Configure it — Run /<skill-name> config to set initial preferences
• Update inventory — only if a publish repo is configured: copy skill to publish repo, add row to SKILLS_GUIDE.md inventory table, add section to README.md. Don't commit.
• Done — All set! Your new skill is ready at ~/.claude/skills/<skill-name>/

Step 10 — Learn from this session

Silently update ~/.claude/skills/skill-creator/preferences.md if the user:

• Renamed the skill (record the rename pattern, e.g. "user prefers verb-led names")
• Removed/added tools you proposed (record the trim/add bias)
• Toggled learning posture against your pre-selection (record the override)
• Cut sections from your draft (record what they consider unnecessary)

Mention once: Noted: you prefer X. Saved for next time.

---

Principles

1. Manifest first. Every generated skill ticks the Structure checklist — no exceptions without an explicit Principles-section justification.
2. Three-tier preferences are the default. Defaults / Profile / Learned, even for small skills. Stateless is the exception, not the rule.
3. Feedback loop is the default. A skill without a /feedback subcommand should be the rare case. Bias toward learning-on in Step 2 Q3.
4. Self-audit is non-negotiable. Step 8 is a checkpoint, not a formality. Don't ship a skill that fails its own audit.
5. Minimal viable skill. Start with the simplest version that works — but "simplest" still includes the learning tier when applicable.
6. Show before write. Always present the planned workflow, knobs, and feedback dimensions before generating the file.
7. Learn and improve. Save patterns from user corrections to make future skill creation faster and more aligned to their style.